The flow panel's figures come from one short-lived cache, so a second /boards load inside 30 s runs none of its reads #7000

A second /boards load inside the window now runs none of the panel's reads. bun x tsc --noEmit, the new test/flow-panel-cache.test.ts, the healthz-gap test, the existing flow-panel* tests and bun scripts/design-lint.ts all pass. I did not run the full check; the forge runs it on this push.

• src/flow/panel_cache.ts (new): it keeps the panel's figures in a WeakMap on the db, as idle_pulls.ts does. A value holds while the write clock (writeStamp) is unchanged, and for 30 s at most.
• Who is cached: the owner's answer and an agent's public-only answer are stored separately. The cache is read only after flowPanel knows the reader, and it is never filled inside a transaction.
• Figures kept: pile, conflicted pulls and rebase lane, requested waits, reviewers busy, idle pulls, parked, lead time, build yield, review latency and openers.
• flow-panel.ts: it draws from panelFigures, so it keeps only the sentence rendering, which still takes now and so stays fresh. The Server-Timing flow phase already wrapped flowPanel in boards-view.ts, so it now times the cached path with no change there.
• Tests in flow-panel-cache.test.ts:

- A second load runs zero queries. - A write makes the next load read again. - The cache expires at 30 s. - An agent's figures are never the owner's: the agent counts 1 opened pull, the owner 2. - The panel renders from the cache.

Left out:

• Not kept in panel_cache.ts: pipeline, throughput and the merge-queue rows still run on every load, because they were not on the issue's list. The held rows keep their own 5 s cache in held.ts, so a second load runs those only once that expires.
• Coarse write clock: it is forge-wide, so any write anywhere ends the window early. That costs hit rate, not correctness.

Closes #6974

A second /boards load inside the window now runs none of the panel's reads. bun x tsc --noEmit, the new test/flow-panel-cache.test.ts, the healthz-gap test, the existing flow-panel* tests and bun scripts/design-lint.ts all pass. I did not run the full check; the forge runs it on this push.

• src/flow/panel_cache.ts (new): it keeps the panel's figures in a WeakMap on the db, as idle_pulls.ts does. A value holds while the write clock (writeStamp) is unchanged, and for 30 s at most.
• Who is cached: the owner's answer and an agent's public-only answer are stored separately. The cache is read only after flowPanel knows the reader, and it is never filled inside a transaction.
• Figures kept: pile, conflicted pulls and rebase lane, requested waits, reviewers busy, idle pulls, parked, lead time, build yield, review latency and openers.
• flow-panel.ts: it draws from panelFigures, so it keeps only the sentence rendering, which still takes now and so stays fresh. The Server-Timing flow phase already wrapped flowPanel in boards-view.ts, so it now times the cached path with no change there.
• Tests in flow-panel-cache.test.ts:

- A second load runs zero queries. - A write makes the next load read again. - The cache expires at 30 s. - An agent's figures are never the owner's: the agent counts 1 opened pull, the owner 2. - The panel renders from the cache.

Left out:

• Not kept in panel_cache.ts: pipeline, throughput and the merge-queue rows still run on every load, because they were not on the issue's list. The held rows keep their own 5 s cache in held.ts, so a second load runs those only once that expires.
• Coarse write clock: it is forge-wide, so any write anywhere ends the window early. That costs hit rate, not correctness.